A wind turbine pitch control method based on quantum optimization and neural differential equations

Abstract

This invention discloses a wind turbine pitch control method based on quantum optimization and neural differential equations, belonging to the field of wind turbine pitch control technology. This method integrates a quantum hybrid optimizer, a neural differential equation (NDE) dynamic model, and an edge-quantum cloud collaborative control architecture. Real-time data is collected through edge devices, and the hidden state of the NDE model is initialized via an encoder network. Latin hypercube sampling of pitch angle combinations is used as input to the NDE model to predict the wind turbine's dynamic response. Based on the prediction results, a total objective function is constructed, fitted with a QUBO matrix, and then the optimal pitch angle is solved by a quantum processing unit through quantum annealing. The quantum cloud platform periodically fine-tunes the model and updates the parameters. This invention achieves millisecond-level real-time control, improves the model's generalization ability and adaptability to complex wind fields, smooths power fluctuations, reduces blade load, and balances wind energy capture efficiency and wind turbine lifespan.

Description

Technical Field

[0001] This invention relates to the field of wind turbine pitch control technology, specifically to a wind turbine pitch control method based on quantum optimization and neural differential equations, which is particularly suitable for real-time optimization, dynamic modeling and edge-quantum cloud collaborative control of wind turbine pitch in complex wind field environments. Background Technology

[0002] The pitch control system of a wind turbine is a core subsystem for wind turbine operation above rated wind speed, and its control performance directly affects the power output stability, structural load, and overall lifespan of the turbine. Current wind turbine pitch control technology mainly suffers from the following shortcomings: (1) Limitations of optimization algorithms: Traditional optimization algorithms (such as particle swarm optimization and sparrow algorithm) require a lot of iterative calculations in multi-objective optimization scenarios, which makes it difficult to meet the real-time requirements of pitch control; and in complex wind fields with drastic wind speed fluctuations, the algorithm converges slowly and has insufficient optimization accuracy, and cannot take into account the multi-objective balance of "maximizing wind energy capture, minimizing blade load, and smoothing power output".

[0003] (2) Insufficient model adaptability: Existing pitch control relies on traditional models such as EBM, LSTM, and SVM, which require accurate physical modeling and have limited generalization ability, making it unable to naturally fit the continuous dynamic process of the wind turbine. In complex scenarios of "wind-turbine-tower" coupling, the model accuracy drops significantly, leading to wind turbine performance loss.

[0004] (3) System architecture defects: Traditional distributed, model predictive control architecture has problems such as high control latency (10-100ms), long optimization cycle (hours), large data transmission volume, and low communication security, making it difficult to adapt to the dynamic changes in the wind farm environment and the real-time control requirements of the wind turbine.

[0005] (4) Poor adaptability to complex wind fields: Traditional pitch control cannot effectively handle the different wind conditions of each blade of the wind turbine. In complex wind fields such as wind shear and turbulence, the power output fluctuates greatly (over 40%) and the blade load is high, which seriously affects the life of the wind turbine and the power quality.

[0006] Therefore, there is an urgent need in this field for a new wind turbine pitch control technology that can more accurately characterize system dynamics and has rapid adaptive optimization capabilities, in order to improve power generation efficiency and operational reliability. Summary of the Invention

[0007] I. Purpose of the Invention The purpose of this invention is to provide a wind turbine adaptive pitch control technology based on quantum hybrid optimization and neural differential equations, which solves the problems of "insufficient real-time performance of multi-objective optimization, weak model generalization ability, poor system architecture performance, and low adaptability to complex wind fields" in existing pitch control, and achieves millisecond-level response, model-free generalization, global optimal control, and adaptation to complex wind fields in wind turbine pitch control.

[0008] II. Technical Solution The technical solution of this invention comprises three parts: a quantum hybrid optimizer, a dynamic model of neural differential equations, and an edge-quantum cloud collaborative control architecture, as detailed below: (a) Quantum Hybrid Optimizer The proposed quantum hybrid optimizer maps the pitch control problem to a QUBO form and employs a classical-quantum hybrid architecture to achieve millisecond-level multi-objective decision-making. The core of the quantum hybrid optimizer consists of two parts: quantum annealing and pitch control optimization, and the classical-quantum hybrid architecture.

[0009] (1) Quantum annealing and pitch control optimization 1. Core Idea: Utilizing the global optimization capability of the quantum annealing algorithm to solve multi-objective optimization problems in pitch control. Quantum annealing, based on the quantum tunneling effect, can efficiently handle combinatorial optimization problems.

[0010] 2. Technical mapping: The pitch angle optimization problem is mapped to a quadratic unconstrained binary optimization model.

[0011] S1: Decision variable coding: Adjust the pitch angle of each blade (in The three blades are discretized into binary variables. For example, the three blades are discretized into binary variables. Quantified into N levels, using Each qubit represents a vector of decision variables. (m is the total number of sub-bits).

[0012] S2: QUBO problem format: Transform the multi-objective problem into a QUBO minimization problem: Minimize:

[0013] in It is a symmetric matrix, and its elements are determined by the weights of the objective function.

[0014] 3. Interface with the Neural Differential Equation Model: The specific element values ​​of the QUBO matrix Q are not fixed in advance, but are dynamically generated by the neural differential equation model (described later) based on the real-time system state. This optimizer itself acts as a general solver, receiving the matrix... And output the optimal solution .

[0015] (2) Classical-Quantum Hybrid Architecture 1. Architecture Overview: A hybrid architecture of "classical processing units + quantum processing units" is adopted to reduce quantum resource requirements and improve practicality.

[0016] 2. Responsibilities of each unit: S1: Classic processing unit: It is responsible for collecting wind data (wind speed, wind direction, turbulence intensity) in real time and performing filtering and feature extraction.

[0017] Transform the physical problem into a QUBO problem: construct the objective function matrix Q and constraint penalty terms based on the wind model and unit parameters.

[0018] S2: Quantum Processing Unit It accepts QUBO problems and solves them using the quantum annealing algorithm. It is compatible with quantum processors such as D-Wave, where the problems need to be embedded in a hardware graph structure (such as the Pegasus topology).

[0019] The optimal qubit configuration is output, and the classical unit is decoded into a pitch angle instruction.

[0020] 3. Architectural advantages: Hybrid design reduces dependence on the number of qubits (only tens to hundreds of qubits are needed) and improves the quality of the solution through classical preprocessing.

[0021] (ii) Dynamic Model of Neural Differential Equations (NDE) By combining neural differential equations (NDEs) as response predictors for the wind turbine dynamic system with a quantum hybrid optimizer (QHME) as the decision-maker, a closed-loop adaptive control system is formed. The NDE implicitly learns the continuous dynamics of the system from the data, while the QHME quickly solves the multi-objective optimization problem based on the NDE's predictions, thereby achieving real-time, adaptive, and global optimization control of the pitch angle.

[0022] (1) Construction and training of neural differential equation models 1. Model Definition: The core model is based on the ordinary differential equations of the divine. Its dynamics are described by the following equations:

[0023] :express The latent state of the system at any given time (such as intrinsic variables like aerodynamic loads and structural stresses).

[0024] :express The input vector at any given time includes, but is not limited to: wind speed, wind direction, turbulence intensity, current pitch angle, generator speed, etc.

[0025] : is a function parameterized by a neural network. These are the network parameters to be trained.

[0026] 2. Training data preparation: Source: The training data comes from high-fidelity computational fluid dynamics simulations and historical SCADA data of wind fields. The dataset should cover various typical operating conditions above the rated wind speed (such as normal turbulence, extreme gusts, shear winds, etc.).

[0027] Sample format: Each sample is a time series segment, including the input I(t) at time t and the corresponding system output (such as the actual blade bending moment, generator power, etc.). The time step can be set according to the data acquisition frequency, for example, 100ms.

[0028] 3. Training methods: S1: Loss function: Mean squared error loss is used. ,in For key system parameters (such as power and load) predicted by the model, These are actual measured or simulated values.

[0029] S2: Gradient Calculation: The adjoint sensitivity method is used for gradient backpropagation. Specifically, the odeint_adjoint function in the torchdiffeq library (Python) is used, which can efficiently calculate the gradient of the loss function with respect to the parameter θ without storing all intermediate states of the forward integration, thus greatly saving memory.

[0030] S3: Optimizer: The Adam optimizer is used, with an initial learning rate set to 1×10 in one specific embodiment. -3 It employs an exponential decay strategy (e.g., decaying to 0.9 times the original value every 10 cycles) to balance the convergence speed in the early stages of training with the stability in the later stages.

[0031] 4. Enhanced model stochasticity: To handle the randomness of wind fields, the model can be extended to neural stochastic differential equations:

[0032] : Represents a Wiener process (Brownian motion), simulating random perturbations.

[0033] : is a neural network parameterized to φ, used to learn the strength of the random term. This extension can significantly improve the model's prediction robustness in uncertain environments such as turbulence.

[0034] 5. Determination of the initial hidden state: To achieve accurate rolling prediction of the NDE model in real-time control, its hidden state needs to be dynamically initialized based on the current system state. This invention achieves this mapping through an encoder network. The encoder uses current and recent historical observation data (including input data) The system response measurement is the input, and the output is the initial hidden state. The encoder network parameter φ and the dynamic network of the NDE The parameters θ are treated as a whole and jointly trained end-to-end using the same adjoint sensitivity method and loss function during the model training phase. After training, φ and θ together constitute the model parameters, which are then distributed to edge devices by the quantum cloud platform.

[0035] (2) Integration and control process with quantum optimizer This step achieves tight coupling between the neural differential equation predictor (NDE) and the quantum optimizer. The core of this step is transforming the continuous dynamic prediction of the NDE into a QUBO problem solvable by the quantum optimizer. The specific closed-loop control flow is as follows: 1. Real-time data input: The controller collects environmental data in real time. .

[0036] 2. Decision space sampling: sampling from all possible combinations of pitch angles (i.e., decision variables) Within the entire search space, a representative set of sample points is selected according to a specific strategy (such as Latin hypercube sampling). This step aims to avoid exhaustive calculations across a large search space.

[0037] 3. NDE Rolling Forecast and Target Value Calculation: Core operation: For each sampled pitch angle combination... Decode it into specific blade angles and real-time data Both inputs are fed into the pre-trained NDE model. The NDE, acting as a simulator, quickly integrates and predicts the dynamic response of the wind turbine after applying the pitch angle combination within a short future time domain (prediction time domain T), outputting a generator power sequence. and blade root bending moment sequence .

[0038] Target value quantization: Based on the prediction sequence of NDE, scalar values ​​representing the three optimization objectives are calculated (the calculation method is consistent with the original document and is omitted here). Thus, for each sample point... A set of corresponding target values ​​were calculated for each. .

[0039] 4. QUBO matrix fitting and quantum solution: S1: Construction and Constraint Handling of the Overall Objective Function: For each sample point, the overall objective function is a linear weighted sum:

[0040] At the same time, pitch angle range constraints and pitch rate constraints are added as penalty terms. Penalty weight , The selection principle is to make its magnitude much larger than the objective function value to ensure the effectiveness of the constraint. After normalizing the power and load data, it can be set... Its value range can be adjusted within [10, 1000].

[0041] S2: QUBO matrix regression fitting (key improvement): The key step of this invention lies in utilizing the sampling point set... and its corresponding total objective function value By using least squares regression, a symmetric matrix Q is fitted, such that the quadratic form... It can best approximate reality. S3: Solution: Input the fitted QUBO matrix Q into the quantum processing unit for solution, and obtain the result. Minimize the optimal binary vector And decoded into the optimal pitch angle combination at the current moment. .

[0042] 5. Control Execution and Feedback: The optimal pitch angle is sent to the pitch servo system for execution. Simultaneously, actual operational data is collected and fed back to the database for periodic online fine-tuning of the NDE model.

[0043] (III) Edge-Quantum Cloud Collaborative Control Architecture (1) Layered design: 1. Edge Device Layer: Deployed on the wind turbine side, it incorporates a lightweight Neural Differential Equation Model (NDE-Lite) and classical optimization algorithms (such as gradient descent). It is responsible for executing millisecond-level (1-5ms) real-time pitch control. Its core task is to directly control the pitch servo system by using the local NDE-Lite model for rapid and continuous local optimization and command generation based on the latest model parameters and optimization weights issued by the quantum cloud.

[0044] 2. Quantum Cloud Platform Layer: Centrally deploys quantum processing units and high-performance classical computing resources. It is responsible for executing long-term (minute- to hour-level) global tasks, including: (a) training and updating the parameters of the complete NDE model based on massive historical and real-time data; (b) running the aforementioned quantum hybrid optimization process to perform global multi-objective optimization analysis and determine the optimal weight coefficients (α, β, γ); and (c) verifying the effectiveness and security of the new model.

[0045] 3. Secure Communication Layer: Quantum key distribution technology is used to ensure secure communication between the edge and the cloud.

[0046] (2) Workflow: 1. Edge device layer autonomy: Edge devices can independently perform millisecond-level real-time control using the currently loaded NDE-Lite model and parameters.

[0047] 2. Data Upload: The edge device packages the collected operational data and uploads it asynchronously to the quantum cloud platform.

[0048] 3. Global optimization in the cloud: The quantum cloud platform integrates all wind turbine data, uses the full version of the NDE model and the quantum hybrid optimizer to perform global analysis and optimization, and updates the model parameters θ and optimization weights α, β, γ.

[0049] 4. Parameter distribution and update: The quantum cloud platform securely distributes the verified new model parameters and weights to each edge device.

[0050] 5. Edge Model Update: Under the premise of ensuring safety, the edge device incrementally updates the local NDE-Lite model and optimization parameters, thereby achieving continuous optimization of the control strategy without interrupting the real-time control loop.

[0051] III. Beneficial Effects Optimizing Real-Time Performance and Global Competency: This invention utilizes a quantum hybrid optimizer to solve the QUBO model for pitch control, fully leveraging the global search capability of quantum annealing. This method effectively overcomes the shortcomings of traditional optimization algorithms, such as being prone to getting trapped in local optima and slow convergence, achieving millisecond-level multi-objective optimization decisions and significantly improving the real-time performance of control.

[0052] Model Adaptation and Prediction Accuracy: By employing Neural Differential Equations (NDEs) as a dynamic predictor, this invention can implicitly learn the continuous dynamics of a wind turbine system from data. This model does not require precise physical modeling, naturally fits the system dynamics, and exhibits stronger robustness to random disturbances such as turbulence, thereby improving prediction accuracy.

[0053] System Architecture Coordination and Efficiency: Based on an edge-quantum cloud collaborative control architecture, this invention achieves a rational allocation of control tasks. The edge side ensures millisecond-level real-time control, while the cloud side is responsible for global optimization and model updates. The two work together to reduce communication load while ensuring both real-time control and global policy optimization.

[0054] Overall control performance improvement: Closed-loop integration of neural differential equation predictors and quantum optimizers enables pitch control to have look-ahead adaptive capabilities. This ultimately manifests in significantly smoothing power fluctuations, effectively reducing blade loads, thereby improving wind energy capture efficiency, extending turbine life, and improving power quality. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that these embodiments are only for explaining the present invention and are not intended to limit it. Those skilled in the art, after understanding the core ideas of the present invention, can make appropriate modifications, all of which should be covered within the protection scope of the present invention.

[0057] Example: A wind turbine pitch control method based on quantum optimization and neural differential equations This embodiment implements the invention on a single wind turbine, and its core lies in achieving the following: Figure 1 The closed-loop control process is shown. The system hardware configuration includes: edge computing devices (such as high-performance industrial IPCs) deployed locally on the wind turbine, pitch servo systems, sensor systems (anemometers, wind vanes, encoders, etc.), and remotely accessed quantum cloud computing services (such as D-Wave quantum computing cloud services).

[0058] S1: System Initialization and Model Pre-training Before the wind turbine is put into operation, the neural differential equation (NDE) model is first trained offline on the quantum cloud platform.

[0059] Training data preparation: High-fidelity computational fluid dynamics (CFD) software was used to simulate various operating conditions at wind speeds above the rated wind speed (e.g., 12 m / s to 25 m / s), including normal turbulence, extreme gusts (e.g., sudden wind speed changes of ±30%), and wind shear. The collected time-series data included: wind speed... ,wind direction turbulence intensity Generator speed Current pitch angle (i=1,2,3) are used as model inputs ; Collect the bending moment at the blade root and generator output power The data was used as the target true value for model prediction. The data sampling frequency was set to 100Hz, generating more than 100,000 time series samples.

[0060] Model Training: On the high-performance servers of the quantum cloud platform, the NDE model was built using the PyTorch framework and the torchdiffeq library. Network Definition This is a fully connected neural network with two hidden layers (128 neurons each) and the activation function SiLU. The adjoint sensitivity method (using the odeint_adjoint function) is used for gradient calculation. The Adam optimizer (initial learning rate 1e-3, decaying by 0.9 every 10 training epochs) is used to minimize the mean squared error loss between the predicted power and the actual load. Training continues until the loss function converges on the validation set.

[0061] Model delivery: After training is complete, the parameters of the trained NDE model will be delivered. and its corresponding encoder network parameters and the initial multi-objective optimization weights (e.g., set to ), The model is packaged, encrypted using quantum key distribution (QKD) technology, and securely sent to the edge computing device on the wind turbine side. The edge device loads the model, becoming a lightweight version of NDE-Lite, ready to perform real-time control.

[0062] S2: Real-time data acquisition and hidden state initialization When the wind turbine is running, edge devices collect environmental and unit data in real time at a frequency of 100Hz, forming the input vector for the current moment. To initiate rolling predictions for the NDE-Lite model, its initial hidden state needs to be determined. The edge device will store historical observation data (including data from the current time and the most recent 0.5 seconds, totaling 5 time steps) for the current time and the last 0.5 seconds. Input the corresponding actual power and load measurements into the loaded encoder network (parameters are...). The encoder network output is the initial hidden state of the NDE-Lite model at the current time step. .

[0063] S3: Decision Space Sampling and NDE Rolling Forecasting The edge device selects K=50 representative sample points from all possible combinations of pitch angles using the Latin hypercube sampling method. Each sample point It is a binary vector representing a discretized combination of the three blade pitch angles.

[0064] For each sample point The edge device decodes it into specific pitch angle commands. .

[0065] The pitch angle command and real-time input Input the NDE-Lite model together. The model is in its current hidden state. Using the initial values, numerical integration (using the fourth-order Runge-Kutta method) is performed over the future prediction time domain T=2 seconds to predict the power sequence after applying this pitch angle combination. and blade root bending moment sequence .

[0066] S4: QUBO Matrix Fitting and Quantum Solution Target value calculation and QUBO fitting: Predicted sequences based on NDE, for each sample point Calculate scalar values ​​for three targets: average power capture. Load fluctuation Power smoothness Then, a linear weighted sum is used to construct the overall objective function. And add a pitch rate constraint penalty term (penalty weight) Next, using these 50 sample points... and its corresponding total objective function value A symmetric matrix is ​​fitted using linear least squares regression. , making the quadratic form It can best approximate reality. .

[0067] Quantum solution: The obtained QUBO matrix This is achieved by mapping and embedding classical processing units within a classical-quantum hybrid architecture into the hardware graph structure of quantum processing units (such as the D-Wave quantum annealing machine). The quantum annealing process is then initiated to perform optimization computation. The quantum annealing machine returns the lowest-energy qubit configuration (i.e., the optimal solution). The classic unit decodes it into the optimal combination of pitch angles at the current moment. .

[0068] S5: Control Execution and Feedback Update Edge devices will execute optimal pitch angle commands. The data is then sent to the pitch servo system for execution. Simultaneously, the actual operational data (including the system response after execution) is cached and asynchronously uploaded to the quantum cloud platform. The quantum cloud platform periodically (e.g., every 6 hours) collects the operational data from all wind turbines, performs online fine-tuning of the complete NDE model, and reruns the global quantum optimization analysis to update the optimal weight parameters. The updated parameters are then securely verified before being sent to the edge devices, enabling continuous self-optimization of the control strategy.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wind turbine pitch control method based on quantum optimization and neural differential equations, characterized in that... This includes the following steps: S1. System initialization and model pre-training steps: On the quantum cloud platform, the NDE model is trained offline using high-fidelity computational fluid dynamics simulation and historical SCADA data of wind fields; After training is complete, the model parameters will be... and encoder network parameters Encrypted data is sent to edge devices. S2. Real-time Data Acquisition and Hidden State Initialization: Edge devices acquire real-time data on wind speed, wind direction, turbulence intensity, generator speed, and current pitch angle to form the input vector. Input current and recent historical observation data into the encoder network to obtain the initial hidden state of the neural differential equation (NDE) model. ; S3. Decision Space Sampling and NDE Rolling Prediction: From all possible pitch angle combinations, K sample points are selected using the Latin hypercube sampling method, where K is an integer greater than 1; each sample point is decoded into a specific pitch angle combination. , and the input vector and initial hidden state —Same as input NDE model; The NDE model uses differential equations The system dynamics are described, and numerical integration is performed over the prediction time domain T to predict the generator power sequence. and blade root bending moment sequence ; S4. QUBO Matrix Fitting and Quantum Solution: Based on the predicted sequence, calculate the average power capture for each sample point. Load fluctuation and power smoothness The target value, and through public Calculate the overall objective function value, and incorporate the pitch angle range constraint and pitch rate constraint as penalty terms. Penalty weight The magnitude of the objective function value is much larger than the objective function value; using the K sample points and their corresponding total objective function values, a symmetric QUBO matrix is ​​fitted using linear least squares regression. , making the quadratic form Approaching ; The QUBO matrix The optimal binary vector is obtained by inputting it into a quantum processing unit and solving the problem using the quantum annealing algorithm. And decoded into the optimal pitch angle combination at the current moment. ; S5. Control Execution and Feedback: The optimal pitch angle combination is sent to the pitch servo system for execution; at the same time, the actual operation data is uploaded to the quantum cloud platform.

2. The method according to claim 1, characterized in that... The NDE model is trained using the adjoint sensitivity method for gradient backpropagation, and the Adam optimizer is used to minimize the mean square error loss between the predicted power and load and the actual values.

3. The method according to claim 1, characterized in that... In step S2, the NDE model is numerically integrated using the fourth-order Runge-Kutta method.

4. The method according to claim 1, characterized in that... In step S3, the penalty weight The range of values ​​is within Within.

5. The method according to claim 1, characterized in that... In step S3, the quantum processing unit is a D-Wave quantum processor, and the QUBO matrix... Before solving, it needs to be mapped and embedded into the processor hardware graph structure by classical processing units.

6. The method according to claim 1, characterized in that... The functions in the NDE model It is a fully connected neural network with two hidden layers, each with 128 neurons, and an activation function of SiLU.

7. The method according to claim 1, characterized in that... In step S4, the quantum cloud platform periodically fine-tunes the NDE model online based on the uploaded running data and reruns the global optimization analysis to update the weight coefficients. Then, the verified new parameters are securely sent to the edge device.

8. The method according to claim 1, characterized in that... The edge device communicates securely with the quantum cloud platform using quantum key distribution technology.

9. A control system for implementing the method according to any one of claims 1-8, characterized in that... The system includes: edge computing devices, pitch servo systems, and sensor systems deployed on the wind turbine side, as well as a quantum cloud computing platform that provides remote access and D-Wave quantum computing cloud services.

Images